Learning VAE with Categorical Labels for Generating Conditional Handwritten Characters

The variational autoencoder (VAE) has succeeded in learning disentangled latent representations from data without supervision. Well disentangled representations can express interpretable semantic value, which is useful for various tasks, including image generation. However, the conventional VAE model is not suitable for data generation with specific category labels because it is challenging to acquire categorical information as latent variables. Therefore, we propose a framework for learning label representations in a VAE by using supervised categorical labels associated with data. Through experiments, we show that this framework is useful for generating data belonging to a specific category. Furthermore, we found that our framework successfully disentangled latent factors from similar data of different classes.